machine learning - Combine training data and validation data, how to select hyper-parameters? -

suppose split data training set , validation set. perform 5-fold cross-validation on training set obtain optimal hyper-parameters model, use optimal hyper-parameters train model , apply resulting model on validation set. question is, reasonable combine training , validation set, , use hyper-parameters obtained training set build final model?

it resonable if training data relatively small , adding validation set makes model stronger. however, @ same time, adding new data makes selected hyperparameters possibly suboptimal (it hard show kind of transformation of hyperparameters should apply when add new data training set). balance 2 things - gain in model quality more data , possible loss due hard predict change in hyperparameters meaning. extent can simulate process make sure makes sense, if have n points in training data , m in validation, can try split training further chunks same proportion (thus 1 n * (n/(n+m) , other n * (m/(n+m))), train on first 1 , check whether optimal hyperparameters transfer (more or less) optimal 1 on whole training set - if so, can safely add validation should transfer well. if not - risk might not worth gain.